Current RMC Research Projects

Background: The National Predictive Services (NPS) has asked the USFS Rocky Mountain Center for FireWeather Intelligence (RMC), a member of the Fire, Fuel and Smoke Science Program at the USFS Rocky Mountain Research Station (RMRS) to develop a system of statistical models for predicting the ignition probability & growth potential of significant fires on a national grid using NWS weather forecast fields as input. This Research is conducted under a Joint Venture Agreement with the CSU CIRA.

1. Developing a National Gridded Model for Probabilistic Forecasting of Cloud-to-Ground Lightning Flashes.

Since 15-20% of wildfire ignitions are due to cloud-to-ground lightning, and lightning-caused fires account for 60% of the area burned annually in the USA, we viewed the ability to precit the chance of lightning flashes/strikes is an important component (and a first step in the development) of a model aimed at foreacsting wildfire ignitions. Guided by this, we derived a system of logistic regression models computing the probabilities of lighning flashes on a 20-km National grid using a compreghensive 3-D dataset of meteorological parameters spanning a period of 27 years extracted from the NOAA North American Regional Reanalysis (NARR). Observed lightnig-flash data were provided by VAISALA NLDN.

The statistical method involves the use of Principal Component Analysis (PCA) with orthogonal rotation to reduce the large cohort of 3-D meteorological driving variables from the NARR dataset to a smaller subset of statistically significant lightning predictors, which were then subjected to a logistic regression analysis to produce equations for estimating the probably of one or more and 10 or more cloud-to-ground (CG) flashes. The resulting set of equations is applicable to forecast fields generated by GFS, WRF or other numerical weather prediction models to generate CG forecast probabilities. The "R" statistical package is utilized to perform PCA and the logistic regressions required to derive the final equations. Resampling the original datasets to a common 20-km resolution national grid was done using the GEMPAK (GEneral Meteorology PAcKage) software jointly developed by NASA and Unidata.

The results is a set of logistic equations for predicting probabilities of CG flashes. Each month of the year is described by 8 logistic equations, i.e. one equation for every 3-hour period of the average 24-hour diurnal cycle for that month (the diurnal averages are computed across 27 years). Furthermore, ConUS is divided into 10 geographical reasons, each one characterised by its own 3-h monthly equations. Hence, the number of monthly predictive equations derived across the Conterminous USA is 8 x 10 = 80. Due to insufficient CG lightning data during winter months, the quarter from December through February is represented by a single diurnal set of 3-h logistic equations. Thus, the total number of predictive logistic equations for all months of the year in all 10 Regions is about 720.

For more details regarding the statistical methods used and the resulting lighning equations please review this 2019 Report.

The lightning prediction algorith is operationally implemented with NWS GFS output to produce lightning forecasts over ConUS twice per day.

2. Developing a National Gridded Model for Probabilistic Forecasting of Wildfire Ignitions.

Three 25-year long grided records of weather reanalysis fields, observed lightning flashes and wildfire occurrences over ConUS were combined to produce daily datasets (from 12:00 to 12:00 UTC) of 20-km horizontal resolution:
a) The NOAA North American Regional Reanalysis (NARR) containing 3-D fields from 00:00, 03:00, 06:00, 09:00, 12:00, 15:00, 18:00, and 21:00 UTC for the period January 1990 - December 2018 were downloaded in GRIB2 format and archived on the RMC machines;
b) U.S lightning data for the lower 48 states were obtained via BLM and gridded at 3-hour increments for the period from January 1990 through December 2018;
c) Wildfire Occurrence Data from 1992 to 2018 provided by Karen Short's database.
d) Monthly Leaf Area Index (LAI) from WRF, NOAA's Evaporative Demand Drought Index (EDDI), digital maps of vegetation cover types and percent vegetation cover;

We employed a statistical method that is similar to the one used in the development of the lightning forecast model above. It involved the application of Principal Component Analysis (PCA) with orthogonal rotation to reduce the initial set of 195 meteorological driving variables from the lower atmosphere of the NARR dataset to a smaller subset of statistically significant fireignition predictors. The reduced set of predictors was then subjected to logistic regression analyses to yield equations for computing the probabilities of one or more wildfire ignitions.The "R" statistical package was utilized to perform PCA and the logistic regressions required to derive the final set of equations. Resampling the original datasets to a common national grid of 20-km horizontal resolution was done using the GEMPAK (GEneral Meteorology PAcKage) software package. We found that monthly predictive equations of ignition probabilities better captured changing environmental conditions during the course of a fire season than a set of seasonal equations. Due to scarcity of wildfire ignitions in the winter and fall, 6 the derivation of monthly logistic equations was limited to the active fire season only. Ignition forecasts for the period from October through December employ equations developed for September, while the May equations are applied to the period January - April. A typical month is described by three equations, one for each wildfire ignition cause: lightning, human, and any cause.

The resulting predictive equations can be applied with output from any numerical weather prediction model as well, such as GFS and WRF to forecast ignition probabilities.

For more details regarding the statistical methods used and the resulting lighning equations please review this 2021 Report.

The wildfire-ignition prediction algorith is operationally implemented with NWS GFS output to produce wildfire-start forecasts over ConUS twice per day.